Universal Monte Carlo Event Generator Nobuo Sato Supported by - - PowerPoint PPT Presentation

▶
universal monte carlo event generator
SMART_READER_LITE
LIVE PREVIEW

Universal Monte Carlo Event Generator Nobuo Sato Supported by - - PowerPoint PPT Presentation

Aug 26, 2022 290 likes •761 views

Universal Monte Carlo Event Generator Nobuo Sato Supported by Jefferson Lab Laboratory CHEP19, Adelaide research and development (LDRD19-13) 1 / 18 Partnership with computer scientists Y. Alanazi (ODU) M. P. Kuchera (Davidson College) Y.

slide-1
SLIDE 1

1 / 18

Universal Monte Carlo Event Generator

Supported by Jefferson Lab Laboratory research and development (LDRD19-13)

Nobuo Sato

CHEP19, Adelaide

slide-2
SLIDE 2

2 / 18

Partnership with computer scientists

  • Y. Alanazi (ODU)
  • M. P. Kuchera (Davidson College)
  • Y. Li (co-PI) (ODU)
  • T. Liu (JLab)
  • R. E. McClellan (JLab)
  • W. Melnitchouk (PI) (JLab)
  • E. Pritchard (Davidson College)
  • R. Ramanujan (Davidson College)
  • M. Robertson (Davidson College)

NS (co-PI) (JLab)

  • R. R. Strauss (Davidson College)
  • L. Velasco (Dallas)
slide-3
SLIDE 3

3 / 18

The big picture

hadrons as emergent phenomena of QCD

quarks and gluons

slide-4
SLIDE 4

3 / 18

The big picture

hadrons as emergent phenomena of QCD

quarks and gluons nucleon structure

slide-5
SLIDE 5

3 / 18

The big picture

hadrons as emergent phenomena of QCD

quarks and gluons nucleon structure hadronization

slide-6
SLIDE 6

4 / 18

Motivations

A new era of nuclear physics has started with the JLab 12 GeV program

slide-7
SLIDE 7

4 / 18

Motivations

A new era of nuclear physics has started with the JLab 12 GeV program New tools based on Machine Learning (ML) to boost the discovery potential are needed

slide-8
SLIDE 8

5 / 18

The goals

slide-9
SLIDE 9

5 / 18

The goals

Build a theory-free MCEG

slide-10
SLIDE 10

5 / 18

The goals

Build a theory-free MCEG Map out particles correlations without biases from approximated theory

slide-11
SLIDE 11

5 / 18

The goals

Build a theory-free MCEG Map out particles correlations without biases from approximated theory MCEG as a data storage utility

slide-12
SLIDE 12

6 / 18

Nature e− P

slide-13
SLIDE 13

6 / 18

Nature

experimental detector

e− P

slide-14
SLIDE 14

6 / 18

Nature

experimental detector detector level events

e− P

slide-15
SLIDE 15

6 / 18

Nature

vertex level events experimental detector detector level events

e− P

slide-16
SLIDE 16

6 / 18

Nature

vertex level events experimental detector detector level events

e− P

slide-17
SLIDE 17

7 / 18

Nature

vertex level events experimental detector detector level events

slide-18
SLIDE 18

7 / 18

Can we use ML to:

Nature

vertex level events experimental detector detector level events

slide-19
SLIDE 19

7 / 18

Can we use ML to: simulate vertex level events?

Nature

vertex level events experimental detector detector level events

slide-20
SLIDE 20

7 / 18

Can we use ML to: simulate vertex level events? simulate detector level events?

Nature

vertex level events experimental detector detector level events

slide-21
SLIDE 21

7 / 18

Can we use ML to: simulate vertex level events? simulate detector level events? simulate nature?

Nature

vertex level events experimental detector detector level events

slide-22
SLIDE 22

7 / 18

Nature

vertex level events experimental detector detector level events

UMCEG

vertex level events detector simulator detector level events

slide-23
SLIDE 23

7 / 18

Nature

vertex level events experimental detector detector level events

UMCEG

vertex level events detector simulator detector level events

datacompression

slide-24
SLIDE 24

8 / 18

Our strategy

Event level ML training → GAN

slide-25
SLIDE 25

8 / 18

Our strategy

Event level ML training → GAN Use a dual GAN as the event generator ρ(particles|multiplicity

  • vectors generator

) × ρ(multiplicity)

  • multiplicity generator
slide-26
SLIDE 26

9 / 18

Challenges

Find optimal data representation → what is the image of an event?

slide-27
SLIDE 27

9 / 18

Challenges

Find optimal data representation → what is the image of an event? How to make the GAN to learn the features of the event? → CNN

slide-28
SLIDE 28

9 / 18

Challenges

Find optimal data representation → what is the image of an event? How to make the GAN to learn the features of the event? → CNN How to escalate from low to higher multiplicities?

slide-29
SLIDE 29

10 / 18

Our current work in progress

Use Pythia as a training and validation tool

slide-30
SLIDE 30

10 / 18

Our current work in progress

Use Pythia as a training and validation tool Ignore detector effects

slide-31
SLIDE 31

10 / 18

Our current work in progress

Use Pythia as a training and validation tool Ignore detector effects Start with inclusive particle generator ρ(particles|multiplicity) → ρ(particles + X)

slide-32
SLIDE 32

11 / 18

Pythia

vertex level events detector level events

UMCEG

vertex level events detector level events

slide-33
SLIDE 33

12 / 18

γ e− e+µ−µ+ νe ¯ νe νµ ¯ νµ ντ ¯ ντ p ¯ p n ¯ n π+π−K+ K− K0

L ¯

K0

L

10−5 10−4 10−3 10−2 10−1 100

probabilities

Pythia GAN

Multiplicity generator

slide-34
SLIDE 34

13 / 18

Generator

FC NN FC NN FC NN

Pythia

l + p → l′ + X Discriminator

FC NN FC NN FC NN

MMD

kx ky kz kikj kT k0 kz/kT kx ky kz kikj kT k0 kz/kT px py pz px py pz

Features Transform

pi → ki pi → ki

Features Extension

kx ky kz kikj kT k0 kz/kT kx ky kz kikj kT k0 kz/kT px py pz px py pz Wasserstein Loss MMD Loss

z ∈ N(0, 1)

Vectors generator

slide-35
SLIDE 35

13 / 18

Generator

FC NN FC NN FC NN

Pythia

l + p → l′ + X Discriminator

FC NN FC NN FC NN

MMD

kx ky kz kikj kT k0 kz/kT kx ky kz kikj kT k0 kz/kT px py pz px py pz

Features Transform

pi → ki pi → ki

Features Extension

kx ky kz kikj kT k0 kz/kT kx ky kz kikj kT k0 kz/kT px py pz px py pz Wasserstein Loss MMD Loss

z ∈ N(0, 1)

Event image = l′

x,y,z

Vectors generator

slide-36
SLIDE 36

13 / 18

Generator

FC NN FC NN FC NN

Pythia

l + p → l′ + X Discriminator

FC NN FC NN FC NN

MMD

kx ky kz kikj kT k0 kz/kT kx ky kz kikj kT k0 kz/kT px py pz px py pz

Features Transform

pi → ki pi → ki

Features Extension

kx ky kz kikj kT k0 kz/kT kx ky kz kikj kT k0 kz/kT px py pz px py pz Wasserstein Loss MMD Loss

z ∈ N(0, 1)

Event image = l′

x,y,z

Feature extension: l′

i · l′ j, l′ 0, l′ z/l′ T

Vectors generator

slide-37
SLIDE 37

13 / 18

Generator

FC NN FC NN FC NN

Pythia

l + p → l′ + X Discriminator

FC NN FC NN FC NN

MMD

kx ky kz kikj kT k0 kz/kT kx ky kz kikj kT k0 kz/kT px py pz px py pz

Features Transform

pi → ki pi → ki

Features Extension

kx ky kz kikj kT k0 kz/kT kx ky kz kikj kT k0 kz/kT px py pz px py pz Wasserstein Loss MMD Loss

z ∈ N(0, 1)

Event image = l′

x,y,z

Feature extension: l′

i · l′ j, l′ 0, l′ z/l′ T

WGAN+MMD

Vectors generator

Butter, Plehn, Winterhalder (’19)

slide-38
SLIDE 38

14 / 18

Validation

slide-39
SLIDE 39

14 / 18

Validation

Relevant observables for inclusive DIS Q2 = −(l − l′)2 xbj =

Q2 2P·(l−l′)

slide-40
SLIDE 40

14 / 18

Validation

Relevant observables for inclusive DIS Q2 = −(l − l′)2 xbj =

Q2 2P·(l−l′)

xbj, Q2 not included as features

slide-41
SLIDE 41

15 / 18

0.01 0.1 1

xbj

10−2 10−1 100 101

Normalized Yield

GAN Pythia

101 102 103 104

Q2 (GeV2)

10−9 10−7 10−5 10−3 10−1

Normalized Yield

GAN Pythia

5 10 15 20 25 30

pT (GeV)

10−5 10−4 10−3 10−2 10−1 100

Normalized Yield

GAN Pythia

Error bands generated with bootstrapped samples

slide-42
SLIDE 42

16 / 18 101 102

Q2

Pythia

0.001 0.01 0.1 1

xbj

101 102

GAN

Isocontours are in agreement xbj, Q2 correlation is learned without adding xbj · Q2 feature

slide-43
SLIDE 43

17 / 18

Summary and outook

It is possible to train a GAN at the event level to build a MCEG

slide-44
SLIDE 44

17 / 18

Summary and outook

It is possible to train a GAN at the event level to build a MCEG The current design provides a blueprint for a generator with higher multiplicity

slide-45
SLIDE 45

18 / 18

Summary and outook

More work is needed, but the results are encouraging

slide-46
SLIDE 46

18 / 18

Summary and outook

More work is needed, but the results are encouraging A fully trained UMCEG will be a complementary tool to theory-based MCEGs such as Pythia